Google Cloud Rapid Agent Hackathon

Observable agents that take approved action

Failed orders --
p95 checkout --
Safety gate Required
Trace status Ready

Current incident

Select a scenario

--

Agent goal

Plan

Waiting

    Tool calls

    Waiting

    Root cause

    Run the agent to investigate.

    Phoenix-style observability

    Trace timeline

    OpenInference spans
    Span Tool Latency Tokens Eval

    Evaluation scores

    Demo run

    Self-review loop

    Phoenix MCP

      Agent Builder contract

      Gemini + Google Cloud runtime

      Cloud Run ready
      Goal Gemini plan Tools Approval Phoenix MCP review

      The repository includes an agent contract that maps the demo to Gemini, Google Cloud Agent Builder, Cloud Run, Firestore, Secret Manager, and Arize Phoenix MCP.

      Safety policy

      Human gated
      • Require at least two evidence sources before proposing a mitigation.
      • Block config rollback, deployment, refund, billing, routing, and customer notification until approval.
      • Record execution and evaluation results as auditable trace evidence.

      Partner MCP role

      Arize track
      • Phoenix MCP reviews traces, spans, and failed evaluation samples.
      • The self-review loop identifies weak reasoning and missing tool calls.
      • Next-run recommendations update prompt, tool, and evaluation criteria.